Use cases

Give your team agents that do real work, on their own, under your rules.

Each page here is a complete story: the problem, what your team gets, how Marshal does it, and the steps to build it yourself. Pick the one that sounds like your week.

Automate team work

Standing jobs that run without anyone pressing a button. A deputy picks them up from an event or a schedule, and a person steps in only when a decision matters.

Every new pull request gets a careful first review within minutes, before a human reviewer opens it.

Review every pull request

Wake up to tested upgrade pull requests. The deputy asks you before it touches a breaking change.

Nightly dependency upgrades

Every alert arrives with a diagnosis, the suspect change and a proposed next step.

Triage alerts from your tools

Give every engineer an agent

Put a capable agent next to every person on the team, in an environment you approved.

Each engineer opens a space with the approved stack, the right credentials and only the network access you allow.

A governed dev environment for every engineer

Hand off many tasks at once. Each run works in its own space, and you review each result on its run page.

Run a fleet of agents in parallel

Stay in control

Agents push code and call your APIs with credentials Marshal injects for them. The agent never holds the token.

Let agents use secrets they never see

Work from anywhere

Hand a task to Marshal from the AI client you already use, and get the result back in the same conversation.

Delegate from your AI client

Most of these stories follow one pattern. Something starts the work, a deputy takes it, and a run does it in its own space. The result lands where your team works, and when the agent needs a decision, it asks a person and waits.

Trigger. A GitHub event, a webhook from your own system, a schedule, or a person. See Schedules and triggers.
Deputy. A standing worker with its own manifest, owner and memory. See Deputies.
Run. One piece of work in a fresh space, with a transcript, a diff and an audit trail. See Runs.
A person. Questions reach Home and, when Slack is connected, a Slack message. A run that waits parks and frees its CPU and memory. See Timeouts and parking.

Before you start

Every story uses the same few building blocks:

A Claude credential for the agent.
A manifest with the agent and the tools the job needs.
For standing work, a deputy and a mission.

More recipes

Short recipes with a prompt you can copy:

On a schedule: Nightly CI triage, Monday engineering digest, Stale PR and issue sweeper, Weekly security check, Cloud cost report.
On an event: Issue triage.
With you in the loop: Approval before merge, On-call helper.